Abstract
The task of single-source domain generalization (SDG) in medical imagesegmentation is crucial due to frequent domain shifts in clinical imagedatasets. To address the challenge of poor generalization across differentdomains, we introduce a Plug-and-Play module for data augmentation calledMoreStyle. MoreStyle diversifies image styles by relaxing low-frequencyconstraints in Fourier space, guiding the image reconstruction network. Withthe help of adversarial learning, MoreStyle further expands the style range andpinpoints the most intricate style combinations within latent features. Tohandle significant style variations, we introduce an uncertainty-weighted loss.This loss emphasizes hard-to-classify pixels resulting only from style shiftswhile mitigating true hard-to-classify pixels in both MoreStyle-generated andoriginal images. Extensive experiments on two widely used benchmarksdemonstrate that the proposed MoreStyle effectively helps to achieve gooddomain generalization ability, and has the potential to further boost theperformance of some state-of-the-art SDG methods.